by Jordan Fulghum, August 2026
I think intelligence is no longer the bottleneck.
This became true somewhere in the past few months, across the most recent frontier model releases and the rapidly encroaching open models. We now have machines that can reason, write, research, analyze, code, and generally knock around various knowledge work at a level that would have felt crazy implausible a couple years ago.
As software engineers, this can be easy to miss.
We tend to reach for the biggest, baddest, most expensive model available. There are good reasons for that. When an agent is modifying a production codebase, debugging a distributed system, or working on its own for an hour+, you want as much brain as you can get, especially if your tokens are heavily subsidized by both your employer and the labs.
As a consequence, though, I think we have developed a blind spot for smaller models.
The small models have gotten extremely capable. More importantly, they are fast and almost comically cheap. As Calvin French-Owen recently pointed out, you can now throw a surprising amount of intelligence at a problem for pennies and most problems do not need a PhD-level brain applied to them. Instead they need inexpensive tokens thrown at the right place, at the right time.
Most normie work is like: ‘read an email, figure out what it means, and then like forward it to the right person’. Turn a meeting into a list of commitments, then check whether those commitments happened. Ask someone for a missing number, move the number from one system to another, and then explain what you did to the people that should know.
It's trivial in effort but non-trivial in importance: companies would fall apart without it.
Even much of the work we consider technical is less sophisticated than the average task we now hand to a coding agent. Most internal software work is some combination of reading structured data, applying known business rules, moving information around, and communicating what happened. The intelligence required to do this is already abundant!
What is missing is everything around it. The model needs access to the inbox, calendar, CRM, ERP, shared drive, and the 36-tab spreadsheet that somehow runs the company. It needs to know who people are, what the business calls things, what it is allowed to do, and when it should stop and ask for help. It needs memory, very specific permissions, an audit trail, and an interface.
Models are still very hobbled by their harness.
We engineers adopt new tools constantly. We install the beta, switch editors annually, or replace our entire workflow because someone on HN said a new terminal is 8% faster. Tinkering is part of the job, and maybe part of our identity. Most people are not like this!
Changing how you do your job is scary, especially when your income depends on doing that job reliably. A workflow that looks inefficient from the outside may contain years of accumulated habit, trust, institutional knowledge, and scar tissue.
Consultants have secretly made billions of dollars over the past twenty years explaining this under the phrase “change management.” People are hard to update.
I've given someone a tool that is objectively faster and they straight up didn't even try it once. I've seen agents that automate 95% of a process and get nowhere because the last 5% does not fit cleanly into the way the company operates. I've spent the better part of a year getting Claude instrumented within a company and watched it sit unused because nobody trusts it, nobody owns it, or using it threatens the status of the person whose cooperation it requires.
Even if we completely unhobble the models tomorrow, the old ways will take years to wash out.
Companies still run on fax machines, shared mailboxes, tribal knowledge, and Excel files named FINAL_v7_ACTUAL_FINAL.xlsx. Obviously better technology exists—even before AI—but it doesn't diffuse automatically.
I think AGI is basically here. You can probably come up with a different definition but I think of AGI as ‘broadly capable machine intelligence that can already perform a huge percentage of economically useful knowledge work’, and it's hard to argue that we're not there already even if it's unevenly distributed.
For better or worse, frontier work will continue but I believe the opportunity now is to flatten that curve.
That means building harnesses, connecting systems, configuring all the tricky permissions and making the agents reliable and boring. Most of this starts with sitting next to the people doing the work and listening, then helping them do their job even better.
The intelligence is there and basically free; it's now our job to help teams catch up.
This post was inspired by a recent conversation with Rylan Schaeffer.
If you are working on this, get in touch.